feat(benchmark): net out retried attempts by trial id in task-level routing stats

[ryan-lempka]

Extends the task-level routing stats (#116) so retried attempts can be netted out of cost, plus a token schema that matches the global stats file.

What changed

1. Retry netting by trial. The proxy sidecar stamps x-switchyard-trial-id (Harbor's trial name, read from the container env) alongside the existing task and session headers. At finalize, requests are grouped by trial, and within a trial the latest attempt is the one Harbor kept (it retries sequentially and discards earlier attempts). So tokens from retried-away attempts land in a retries section and can be subtracted. No timestamps, no result.json parsing.
2. Six-field token schema. Records and the rollup now carry cached_tokens, cache_creation_tokens, and reasoning_tokens alongside prompt/completion/total, matching routing_stats_final.json so the existing cost estimator prices either file. model is the served model id, so buckets reconcile with the global file.

Schema, before → after (per task in routing_stats_by_task.json):

// before
"build-system-task-ordering": {
  "requests": 24,
  "sessions": ["8c54..."],
  "models": [
    {"model": "gpt-5.4-mini", "tier": "strong", "calls": 12,
     "prompt_tokens": 349072, "completion_tokens": 22239, "total_tokens": 371311}
  ]
}

// after
"build-system-task-ordering": {
  "requests": 29,
  "n_retries": 1,
  "final": {
    "calls": 24,
    "totals": {"prompt_tokens": 349072, "cached_tokens": 300112, "cache_creation_tokens": 0,
               "completion_tokens": 22239, "reasoning_tokens": 18110, "total_tokens": 371311},
    "models": [
      {"model": "gpt-5.4-mini", "tier": "strong", "calls": 12,
       "prompt_tokens": 349072, "cached_tokens": 300112, "cache_creation_tokens": 0,
       "completion_tokens": 22239, "reasoning_tokens": 18110, "total_tokens": 371311}
    ]
  },
  "retries": { "calls": 5, "totals": { /* same six fields */ }, "models": [ /* ... */ ] }
}

final and retries share the same shape, so the cost estimator runs on either: net cost = price(final), retry overhead = price(retries).

Tested end to end: Ran a real Harbor benchmark, forced a task to time out and retry, and confirmed the stamped trial id matches Harbor's trial name and that the retried attempt's tokens land in retries while the kept attempt's stay in final.

[coderabbitai]

Walkthrough

Changes

Routing accounting

Layer / File(s)	Summary
Detailed routing records switchyard/lib/processors/routing_log_response_processor.py, tests/test_routing_log_response_processor.py	Routing records now prefer the backend-served model and capture prompt, cache, completion, reasoning, and total token fields across supported provider usage shapes.
Final and retry manifest rollup benchmark/run_manifest.py, tests/test_run_manifest.py, benchmark/README.md	Manifest finalization classifies requests using Harbor trial windows, aggregates final/retry statistics, updates tests, and documents the resulting artifacts.

Estimated code review effort: 4 (Complex) | ~45 minutes

Poem

A rabbit logs tokens, both fluffy and bright,
With models served true and retries in sight.
Final hops stay final, old hops gently retire,
Cache crumbs and thoughts join the routing wire.
Hop, hop—clean stats climb higher!

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title accurately describes the main change: task-level routing stats now separate and net out retries by trial window.

---

_{Comment @coderabbitai help to get the list of available commands.}

[ayushag-nv]

Looks good. Aligned with what I was trying to do with logging per request routing decisions